Skip to lesson
Exit
AI-Editor Systems & Design1 / 2

1 min lesson

Failure handling

Consider this situation: "An agent proposes a five-file refactor. List the failure modes you'd design for before letting it apply and one mitigation each." Start with the decision, then the evidence.

Step 1 of 2

Failure handling

Edits fail in specific, recurring ways and naming them is what separates a careful systems answer from a hopeful one.

Failure
Partial application
What goes wrong
Edit applies to 3 of 5 files, then a tool errors
Recovery
Atomic-per-turn checkpoint; roll back to a consistent state, surface what landed
Failure
Conflicting edit
What goes wrong
User typed in a region the agent is rewriting
Recovery
Detect on apply, re-anchor the hunk or re-generate against the new buffer
Failure
Stale anchor
What goes wrong
The diff targets line numbers that have shifted
Recovery
Anchor edits to content/AST nodes, not raw line offsets
Failure
Doom loop
What goes wrong
Agent retries the same broken fix forever
Recovery
Cap iterations, detect repetition, hand back to the user with what it learned
Interview move

When you describe an agent, always say how it stays in control: the diff is the contract, edits are a transaction, the loop verifies before it stops. Cursor's bar is "powerful tools without compromising ease-of-use," so an answer that pairs capability with a rollback story and a review surface lands far better than raw autonomy.